Generative modelling has recently emerged as a promising tool to efficientlyexplore the vast chemical space. In homogeneous catalysis, Transition MetalComplexes (TMCs) are ubiquitous, and finding better TMC catalysts is critical to anumber of technologically relevant reactions. Evaluating reaction rates requiresexpensive transition state (TS) structure search, making traditional library-basedscreening difficult. Inverse-design of TMCs with a model capable of generatinggood TS guesses can lead to breakthroughs in catalytic science. We present suchgenerative model herein. The model is an instance of an equivariant conditionaldiffusion model, and the key innovation lies in its specific data representation andtraining procedure, that allow generic databases (e.g. non-TS structures) to beleveraged at training time, while offering the desired controllability at samplingtime (e.g. ability to generate TSs on demand). We demonstrate that augmentingthe training database with generic (but related) data enables a practical level ofperformance to be reached. In a case study, our model successfully explores thechemical space around Vaska’s complex, where the property of interest is theH2-activation barrier, in two distinct settings: generation from scratch, and redesignof a specific ligand in a known TMC. In both cases, we validate a selection of novelsamples with Density Functional Theory (DFT) calculations